Translation: Vernacular Blockchain

For many, the future AI agent is like the character J.A.R.V.I.S. from the Marvel Cinematic Universe.

J.A.R.V.I.S., which stands for "Just A Rather Very Intelligent System," was originally a natural language computing system created by Tony Stark—a fictional industrialist and genius investor. Later, it became an AI system serving as Stark's assistant. Eventually, J.A.R.V.I.S. gained a synthetic body and transformed into the robot 'Vision.'

Although AI agents—autonomous and semi-autonomous generative AI systems capable of acting independently—may be far from having physical capabilities, they could come close to or even surpass J.A.R.V.I.S. at some point next year.

The growing popularity of AI agents in the second half of 2024 resembles the rapid rise of ChatGPT and other generative AI systems in 2022, which transformed the AI market. Vendors seem to be quickly shifting from developing the latest large language models (LLMs) and AI chatbots to creating agents and action models.

For example, Salesforce launched Agentforce, a low-code agent building tool, last fall. Microsoft launched the AI Agents Service, a community platform to help developers build AI agents.

Other vendors have also introduced AI agents into businesses, automating various business processes. Forrester Research listed 400 vendors currently building agents.

"Forrester Research analyst Craig Le Clair said: 'The excitement around them is very high right now.' He said, 'But there are also certain risks, as you are unleashing an automation process that can be executed automatically without human oversight and balance.'"

Excitement and risk coexist, meaning that AI experts and vendors have high expectations for AI agents in 2025.

1. Eliminate confusion through real applications

One expectation is that while 2024 lays the groundwork for AI agents, 2025 will be a year of readiness for enterprises to embrace AI agents, AI market experts say.

This means that the confusion surrounding agents will dissipate, said AJ Sunder, co-founder and CIO of Responsive, an AI-driven proposal and response software vendor.

"There is a lot of confusion between agents and automation, and agents and RPA (robotic process automation)," Sunder said. "Much of this confusion will dissipate. Then we will start to see more agents deployed and used in the real world."

RPA uses robots or automation to automate repetitive tasks without relying on AI, while agents involve AI technologies. RPA is deterministic and predictable, while agents are not.

"What they have in common is that they are both digital colleagues," Le Clair said. "It's just that when you add AI to digital colleagues, we call it an AI agent; it is smarter, understands context, and knows how to avoid getting stuck."

Sunder said that some practical applications of agents will appear in customer service; others will emerge in finance or fraud detection.

"Any complex task requires AI to remember, plan, and execute multi-step, intricate tasks, and I believe agents will play a huge role in this," Sunder said.

One complex application is video creation.

"Many of these agent AI solutions can actually be deployed in a way that assists the video creation process," said Shahzaib Aslam, research director at Colossyan, an AI video platform.

AI agents can help create engaging videos, provide compelling arguments, and include calls to action that encourage customers to take actions like purchasing products, Aslam said.

"This becomes a very powerful tool because it will help you create a video that is more engaging and has a higher success rate," he said.

Agents will not only play a role in different application scenarios like video creation, but many will also start using them to solve scaling issues, said Gartner analyst Tom Coshow.

However, the applications and uses of AI agents come at different levels, said Peter van der Putten, Director of AI Labs and Chief Scientist at Pegasystems, a workflow automation and decisioning vendor.

He stated that at one extreme of application, AI agents can read, integrate, and synthesize information to draw certain conclusions but will take no action. At the other end, AI agents take action based on the information they synthesize.

"The true success of agents does not lie in the intelligence capabilities of these agents themselves, but in how they are embedded into real applications," he said.

However, he continued to state that most businesses must first experiment before seeing the value of AI agents.

"Sometimes I am even surprised by what these systems can do," van der Putten said. "The only way to understand this is through safe experimentation."

2. Better reasoning models

Another expectation regarding AI agents is that large language models (LLMs) will continue to serve as their brains. This means LLMs need to become stronger in reasoning so that AI agents can perform their tasks better.

Aslam said that chain of thought prompting has already demonstrated this.

The concept is that models generate not just one answer to a question, but multiple answers, reasoning through a series of steps to ultimately arrive at a conclusion.

Although this may be costly, as enterprises need to run multiple reasoning tasks to generate chain of thought, it also enhances the reasoning capabilities of the models, Aslam said.

He added that this will be a field that the AI industry and academia will delve into in 2025.

"This way of adding explainability into models makes a lot of sense, and we will see more work and research focused on this direction, that is, expanding computational scale during reasoning and enabling models to make predictions in a systematic and reasoning way, rather than just generating content simply," he continued.

3. Specific task agents

While more agent-based application scenarios may emerge by 2025, it will not eliminate the need for human intervention.

However, the fear that jobs will be replaced persists as AI agents bring new levels of automation.

Some in the industry indicate that while AI agents will have a certain level of autonomy by 2025, they will not be fully autonomous. In other words, AI agents will perform part of a person's job but will not take over the entire role. For example, an AI agent might help you find the contact information for a travel agency you want to use, but it cannot complete the entire booking process.

"We will see agents not completely independently taking over entire jobs, but rather taking on part of a person's responsibilities or part of a process, and then working alongside traditional automation systems, human collaboration, and other agents," said Mark Greene, Senior Vice President and General Manager at UiPath.

Agents that take on part of the responsibility will be specialized and perform tasks in a singular manner. This will make AI agents more precise in accomplishing tasks, Greene said.

"The clearer the responsibility, the easier it is to measure its effectiveness," he said.

4. Infrastructure for AI agents

In addition to the rise of single-task AI agents, 2025 may also become a year for building AI agent infrastructure, said Futurum Group analyst Olivier Blanchard.

To enable AI agents to communicate with other agents and even collaborate with humans to perform tasks, a coordination layer is needed, Blanchard said.

"2025 will not be the year we see fully mature agent-based AI," he said. "2025 is the year we build the infrastructure for it, the foundational framework to construct it."

He added that vendors who may help build this infrastructure include chip manufacturers like Qualcomm, Intel, and AMD.

"Qualcomm's processors will primarily be used for agent-based AI on devices," Blanchard continued. Meanwhile, Nvidia's processors are currently more focused on cloud collaboration with agent-based AI.

"Nvidia's GPUs have already been widely used to train AI models, which lays the foundation for future agent-based AI layers," he said. "In two years, agent-based AI will be a mix of cloud and device software, working together."

Currently, Nvidia is primarily collaborating with the cloud, while Qualcomm is focusing on devices. On the other hand, manufacturers like Apple and Samsung will be involved in creating a coordination layer that enables agent-based AI to work collaboratively across platforms, devices, and applications, Blanchard said.

"We already have these foundations in place," Blanchard said. "What we lack is a system that can 'do everything.'"

5. One way to approach the coordination layer is through multimodal AI.

While generative AI systems like ChatGPT have input-output capabilities, they cannot connect humans to other applications.

However, with the development and maturity of multimodal AI that can generate video output from image input, this will facilitate agent-based AI working better.

"As models become smarter, our agents will also become smarter," Coshow said.

Blanchard said that AI agents need a coordination layer that can work across platforms and devices. The coordination layer consists of links that allow AI agents to switch from one platform or interface to another, or from one application to the next.

If Qualcomm builds its own coordination layer and AMD builds its own coordination layer, this will make interoperability of agent-based AI a major challenge.

"If all chip manufacturers are using their own coordination layers, they may not communicate well with each other," Blanchard said.

6. Challenges faced by agent-based AI in 2025

Like other AI technologies, agent-based AI will face a series of challenges by 2025. One of them is data issues.

Providing the data needed for agent-based AI to execute tasks can become very challenging due to data often being scattered across different sources and processes, Greene stated.

Another issue is the lack of knowledge about agent automation design processes, Greene added.

For instance, the industry needs to understand when humans should interact with agent-based AI, how to interact, and through which channels to communicate with agent-based AI, he said.

Another challenge is the trust issue, Sunder stated.

"If the underlying technology still relies on generative AI and large language models, then these shortcomings will also be inherited by agent-based AI," he said.

Despite these obstacles, Sunder believes that 2025 will be an important year for agent-based AI.

"We will figure out in which scenarios agent-based AI makes sense, how to deploy them, how to earn trust, and then fully let go," he said. "The promise of it being completely autonomous, I believe it will ultimately be achieved; but whether it will be realized by 2025, I think it is unlikely."